def plot(model_dir="experiments/S_BGCN_HoustonDataset_k2_2021_04_26_2"):
gt = load_gt()
dataset = HoustonDatasetMini()
alpha, vac, dis = [unflatten_array(array) for array in load_alpha_vac_dis(model_dir)]
# alpha = unflatten_array(np.load(os.path.join(model_dir, "prob.npy")), gt.shape)
# fig, ax = dense_fig(3, 1, width=7, height=4, keep_square=False)
fig, ax = dense_fig(3, 2, width=8, height=7, keep_square=False)
ax[0, 1].axis('off') # invisible
classification = np.argmax(alpha, axis=-1)
gt_rgb = classes_array_to_colormapped_array(gt)
classification_rgb = classes_array_to_colormapped_array(classification)
ax[0, 0].imshow(np.rollaxis(rgb.training_array(), 0, 3)[:, 1800:3600])
legend = shade_one_axes_with_splits(ax[0, 0], dataset.mask_tr, dataset.mask_va, dataset.mask_te)
ax[0, 0].legend(handles=legend, loc='center left', bbox_to_anchor=(1, 0.5))
ax[1, 0].imshow(gt_rgb)
ax[2, 0].imshow(classification_rgb)
img = ax[1, 1].matshow(vac)
add_colorbar(fig, img, ax[1, 1], x_shift=.1)
img = ax[2, 1].matshow(dis)
add_colorbar(fig, img, ax[2, 1], x_shift=.1)
set_each_ax_y_label(["Optical", "GT", "S-BMLP\nClassification", "Vacuity", "Dissonance"],
[ax[0, 0], ax[1, 0], ax[2, 0], ax[1, 1], ax[2, 1]])
# set_each_ax_y_label(["Optical", "GT", "GCN\nClassification"], [ax[0, 0], ax[1, 0], ax[1, 1]])
remove_ax_list_ticks(ax.flatten())
fig.savefig("docs/25_may/sbmlp.pdf", bbox_inches="tight")
def download(self):
gt = ground_truth.data.array[0]
y = gt[:, 1800:3600].reshape(-1, 1)
x = np.concatenate([rgb.training_array(), hyperspectral.training_array(), lidar.training_array()])
x = x[:, :, 1800:3600]
x = np.rollaxis(x.reshape(x.shape[0], -1), 0, 2)
print("Computing k neighbors graph...")
a = kneighbors_graph(x, 15, include_self=False)
a = a + a.T # to make graph symmetric (using k neighbours in "either" rather than "mutual" mode)
a[a > 1] = 1 # get rid of any edges we just made double
print("Graph computed.")
# Create the directory
os.mkdir(self.path)
filename = os.path.join(self.path, f'graph')
np.savez(filename, x=x, a=a, y=OneHotEncoder().fit_transform(y).toarray())
def channels_figure(margin=.1, width=.6):
# image = [np.arange(50 * 100).reshape(50, 100)] * 4
# shape: channels, height, width
image = np.concatenate([rgb.training_array(), hyperspectral.training_array(), lidar.training_array()])[...,
1800:3600]
aspect, numb = image.shape[1] / image.shape[2], len(image)
height = width * aspect
dx, dy = (1 - 2 * margin - width) / (numb - 1), (1 - 2 * margin - height) / (numb - 1)
fig = plt.figure(figsize=(8, 4))
axes = [fig.add_axes([margin + n * dx, margin + (numb - n - 1) * dy, width, height]) for n in
range(numb - 1 - 2, -1, -1)] # extra -2 because we'll show 3 channels in one for the rgb
for i, ax in enumerate(axes):
if i != len(axes) - 1:
ax.matshow(image[len(axes) - i - 1], cmap='hsv')
else:
ax.imshow(np.rollaxis(image[:3], 0, 3))
ax.axis('off')
fig.savefig('docs/19_oct/channels.pdf', bbox_inches='tight')
def data_figure_rgb():
fig, ax = dense_fig(1, 1)
ax.imshow(np.rollaxis(rgb.training_array()[..., 1800:3600], 0, 3))
fig.savefig("docs/19_oct/data_figure_rgb.pdf", bbox_inches="tight")